Financial Trading

Every microsecond on NYSE and Binance, dozens of algorithms fight for the same order. The loser loses money.

• **Binance** processes up to 1.4 million orders per second and supports WebSocket streams with <10 ms latency for 2,000+ trading pairs at once
• **NYSE** moves up to 5B messages per day over XDP (eXtreme Data Platform); a direct data feed runs $50,000+ per month for institutional clients
• **Virtu Financial** is the HFT firm that traded profitably 1,237 of 1,238 trading days in a row (2009-2014) using co-location and FIX with 50-100 us latency
• **Jump Trading** built a microwave link Chicago-NJ in 2010 for $14M, cutting latency from 13 to 8.1 ms; it paid off in months

Order Book

An order book is a two-sided queue of buy orders (bids) and sell orders (asks) for a financial instrument, sorted by price. Binance handles up to 100,000 order-book updates per second on the BTC/USDT pair during volatility. Each update carries a price level and the new volume at that level.

The order book lives in two modes: full snapshot (complete state, 5-20 MB JSON for top pairs) and incremental updates (deltas, 50-200 bytes). On connect a client gets a snapshot, then subscribes to the delta stream. Missing even one delta makes the local copy invalid, so a fresh snapshot is required.

• **Best bid** - the highest buy price; **best ask** - the lowest sell price
• **Spread** = best ask - best bid; on Binance BTC/USDT it is usually $0.01-$0.10
• **Depth** - total volume across N levels; used to gauge liquidity
• Quantity=0 in a delta means the level is removed (do not skip it!)

Market Data

A market data feed is the stream of trading activity: trades, prices, volumes, indicators. NYSE pushes up to 5 billion messages per day through XDP (eXtreme Data Platform). The data ships via UDP multicast: one stream to thousands of subscribers at once without the TCP overhead per connection.

Market data has two classes: Level 1 (best bid/ask and last trade only) and Level 2 (full order book to N levels). Bloomberg Terminal delivers Level 1 for 35M instruments worldwide. Level 2 for one exchange is already gigabytes per second at peak, requiring co-located servers next to the exchange.

• **Consolidated tape** - aggregated stream across every venue (SIP in the US); up to 500 us latency
• **Direct feed** - direct connection to the exchange; 50-200 us latency, more expensive
• **Snapshot vs stream**: a periodic snapshot is simpler but misses intra-second moves
• Store timestamps in nanoseconds - microseconds occasionally collide on NYSE

FIX Protocol

FIX (Financial Information eXchange) is the standard protocol for electronic trading, created in 1992 by Fidelity and Salomon Brothers. Today 90%+ of institutional trading instructions move over FIX. The protocol is text-based: fields are separated by the SOH character (0x01), which simplifies parsing but inflates message size compared to binary formats.

FIX/FAST (Fast Application-layer Encoding) is the binary FIX extension for market data with compression of repeating fields. CME Group uses FIX/FAST for MDP 3.0 (Market Data Platform), delivering updates with 3-8 us latency on co-located servers. Order routing (sending orders) uses classic text FIX or proprietary binary protocols like OUCH (NASDAQ).

1. **Session layer**: Logon (A), Heartbeat (0), Logout (5) - connection maintenance
2. **Application layer**: NewOrderSingle (D), ExecutionReport (8), OrderCancelRequest (F)
3. **Checksum**: the last field (tag 10) is the sum of all message bytes mod 256
4. **CompIDs**: SenderCompID (49) and TargetCompID (56) identify the parties
5. FIX/FAST compresses a typical 200-byte message down to 20-40 bytes through delta encoding

Trading Latency

In algorithmic trading, latency is a competitive edge. HFT firm Virtu Financial earns on the spread by reacting to market changes faster than anyone else. Co-location in the exchange's data center gives 50-200 us from event to responsive order. Over the public internet it is 5-50 ms, which makes competing with HFT impossible.

• **Co-location**: a server physically inside the exchange data center; lease from $5,000/mo on NYSE
• **Kernel bypass**: DPDK, io_uring - bypass the OS to talk directly to the NIC
• **FPGA**: Field-Programmable Gate Array for market-data parsing in hardware; <1 us
• **Fiber vs microwave**: Chicago-NJ over microwave 8.1 ms vs 13.1 ms over fiber - HFT uses microwave
• **Jitter matters more than the mean**: P99 latency, not the average, decides real losses

Takeaways

• **Order book** is the primary source of truth on liquidity; the client maintains a local copy through snapshot + delta updates with sequence numbers
• **FIX protocol** is the industry standard (1992) for trading instructions; text format with SOH separators; FIX/FAST is the binary extension for market data with up to 5x compression
• **Trading latency** is measured in microseconds; co-location, kernel bypass (DPDK), FPGA, and CPU pinning form the standard HFT stack; jitter hurts more than mean latency

Related topics

Financial trading intersects several real-time-system areas:

• WebSocket internals — Binance and Coinbase deliver market data over WebSocket; understanding framing and backpressure is critical
• Real-time rate limiting — Exchanges throttle API requests (Binance: 1,200 weight/min); a violation triggers a 24h IP ban
• WebSocket security — Trading APIs require an HMAC signature on every request; key compromise means loss of funds

Вопросы для размышления

• Binance updates the order book 100 times per second. A client with slow internet drops some deltas. How do you detect the gap and recover consistency?
• FIX is text and takes 200 bytes per order. Binance uses its own binary WebSocket API. In what scenarios does FIX still win despite the inefficiency?
• An HFT firm spends $14M on a microwave link to save 5 ms. What daily trading volume and spread would pay that back in a year?

Связанные уроки

• db-19-redis